xref: /linux/kernel/rcu/tree_stall.h (revision d6a5c562214f26e442c8ec3ff1e28e16675d1bcf)
1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3  * RCU CPU stall warnings for normal RCU grace periods
4  *
5  * Copyright IBM Corporation, 2019
6  *
7  * Author: Paul E. McKenney <paulmck@linux.ibm.com>
8  */
9 
10 #include <linux/kvm_para.h>
11 
12 //////////////////////////////////////////////////////////////////////////////
13 //
14 // Controlling CPU stall warnings, including delay calculation.
15 
16 /* panic() on RCU Stall sysctl. */
17 int sysctl_panic_on_rcu_stall __read_mostly;
18 int sysctl_max_rcu_stall_to_panic __read_mostly;
19 
20 #ifdef CONFIG_PROVE_RCU
21 #define RCU_STALL_DELAY_DELTA		(5 * HZ)
22 #else
23 #define RCU_STALL_DELAY_DELTA		0
24 #endif
25 #define RCU_STALL_MIGHT_DIV		8
26 #define RCU_STALL_MIGHT_MIN		(2 * HZ)
27 
28 int rcu_exp_jiffies_till_stall_check(void)
29 {
30 	int cpu_stall_timeout = READ_ONCE(rcu_exp_cpu_stall_timeout);
31 	int exp_stall_delay_delta = 0;
32 	int till_stall_check;
33 
34 	// Zero says to use rcu_cpu_stall_timeout, but in milliseconds.
35 	if (!cpu_stall_timeout)
36 		cpu_stall_timeout = jiffies_to_msecs(rcu_jiffies_till_stall_check());
37 
38 	// Limit check must be consistent with the Kconfig limits for
39 	// CONFIG_RCU_EXP_CPU_STALL_TIMEOUT, so check the allowed range.
40 	// The minimum clamped value is "2UL", because at least one full
41 	// tick has to be guaranteed.
42 	till_stall_check = clamp(msecs_to_jiffies(cpu_stall_timeout), 2UL, 21UL * HZ);
43 
44 	if (cpu_stall_timeout && jiffies_to_msecs(till_stall_check) != cpu_stall_timeout)
45 		WRITE_ONCE(rcu_exp_cpu_stall_timeout, jiffies_to_msecs(till_stall_check));
46 
47 #ifdef CONFIG_PROVE_RCU
48 	/* Add extra ~25% out of till_stall_check. */
49 	exp_stall_delay_delta = ((till_stall_check * 25) / 100) + 1;
50 #endif
51 
52 	return till_stall_check + exp_stall_delay_delta;
53 }
54 EXPORT_SYMBOL_GPL(rcu_exp_jiffies_till_stall_check);
55 
56 /* Limit-check stall timeouts specified at boottime and runtime. */
57 int rcu_jiffies_till_stall_check(void)
58 {
59 	int till_stall_check = READ_ONCE(rcu_cpu_stall_timeout);
60 
61 	/*
62 	 * Limit check must be consistent with the Kconfig limits
63 	 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
64 	 */
65 	if (till_stall_check < 3) {
66 		WRITE_ONCE(rcu_cpu_stall_timeout, 3);
67 		till_stall_check = 3;
68 	} else if (till_stall_check > 300) {
69 		WRITE_ONCE(rcu_cpu_stall_timeout, 300);
70 		till_stall_check = 300;
71 	}
72 	return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
73 }
74 EXPORT_SYMBOL_GPL(rcu_jiffies_till_stall_check);
75 
76 /**
77  * rcu_gp_might_be_stalled - Is it likely that the grace period is stalled?
78  *
79  * Returns @true if the current grace period is sufficiently old that
80  * it is reasonable to assume that it might be stalled.  This can be
81  * useful when deciding whether to allocate memory to enable RCU-mediated
82  * freeing on the one hand or just invoking synchronize_rcu() on the other.
83  * The latter is preferable when the grace period is stalled.
84  *
85  * Note that sampling of the .gp_start and .gp_seq fields must be done
86  * carefully to avoid false positives at the beginnings and ends of
87  * grace periods.
88  */
89 bool rcu_gp_might_be_stalled(void)
90 {
91 	unsigned long d = rcu_jiffies_till_stall_check() / RCU_STALL_MIGHT_DIV;
92 	unsigned long j = jiffies;
93 
94 	if (d < RCU_STALL_MIGHT_MIN)
95 		d = RCU_STALL_MIGHT_MIN;
96 	smp_mb(); // jiffies before .gp_seq to avoid false positives.
97 	if (!rcu_gp_in_progress())
98 		return false;
99 	// Long delays at this point avoids false positive, but a delay
100 	// of ULONG_MAX/4 jiffies voids your no-false-positive warranty.
101 	smp_mb(); // .gp_seq before second .gp_start
102 	// And ditto here.
103 	return !time_before(j, READ_ONCE(rcu_state.gp_start) + d);
104 }
105 
106 /* Don't do RCU CPU stall warnings during long sysrq printouts. */
107 void rcu_sysrq_start(void)
108 {
109 	if (!rcu_cpu_stall_suppress)
110 		rcu_cpu_stall_suppress = 2;
111 }
112 
113 void rcu_sysrq_end(void)
114 {
115 	if (rcu_cpu_stall_suppress == 2)
116 		rcu_cpu_stall_suppress = 0;
117 }
118 
119 /* Don't print RCU CPU stall warnings during a kernel panic. */
120 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
121 {
122 	rcu_cpu_stall_suppress = 1;
123 	return NOTIFY_DONE;
124 }
125 
126 static struct notifier_block rcu_panic_block = {
127 	.notifier_call = rcu_panic,
128 };
129 
130 static int __init check_cpu_stall_init(void)
131 {
132 	atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
133 	return 0;
134 }
135 early_initcall(check_cpu_stall_init);
136 
137 /* If so specified via sysctl, panic, yielding cleaner stall-warning output. */
138 static void panic_on_rcu_stall(void)
139 {
140 	static int cpu_stall;
141 
142 	if (++cpu_stall < sysctl_max_rcu_stall_to_panic)
143 		return;
144 
145 	if (sysctl_panic_on_rcu_stall)
146 		panic("RCU Stall\n");
147 }
148 
149 /**
150  * rcu_cpu_stall_reset - restart stall-warning timeout for current grace period
151  *
152  * The caller must disable hard irqs.
153  */
154 void rcu_cpu_stall_reset(void)
155 {
156 	WRITE_ONCE(rcu_state.jiffies_stall,
157 		   jiffies + rcu_jiffies_till_stall_check());
158 }
159 
160 //////////////////////////////////////////////////////////////////////////////
161 //
162 // Interaction with RCU grace periods
163 
164 /* Start of new grace period, so record stall time (and forcing times). */
165 static void record_gp_stall_check_time(void)
166 {
167 	unsigned long j = jiffies;
168 	unsigned long j1;
169 
170 	WRITE_ONCE(rcu_state.gp_start, j);
171 	j1 = rcu_jiffies_till_stall_check();
172 	smp_mb(); // ->gp_start before ->jiffies_stall and caller's ->gp_seq.
173 	WRITE_ONCE(rcu_state.jiffies_stall, j + j1);
174 	rcu_state.jiffies_resched = j + j1 / 2;
175 	rcu_state.n_force_qs_gpstart = READ_ONCE(rcu_state.n_force_qs);
176 }
177 
178 /* Zero ->ticks_this_gp and snapshot the number of RCU softirq handlers. */
179 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
180 {
181 	rdp->ticks_this_gp = 0;
182 	rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
183 	WRITE_ONCE(rdp->last_fqs_resched, jiffies);
184 }
185 
186 /*
187  * If too much time has passed in the current grace period, and if
188  * so configured, go kick the relevant kthreads.
189  */
190 static void rcu_stall_kick_kthreads(void)
191 {
192 	unsigned long j;
193 
194 	if (!READ_ONCE(rcu_kick_kthreads))
195 		return;
196 	j = READ_ONCE(rcu_state.jiffies_kick_kthreads);
197 	if (time_after(jiffies, j) && rcu_state.gp_kthread &&
198 	    (rcu_gp_in_progress() || READ_ONCE(rcu_state.gp_flags))) {
199 		WARN_ONCE(1, "Kicking %s grace-period kthread\n",
200 			  rcu_state.name);
201 		rcu_ftrace_dump(DUMP_ALL);
202 		wake_up_process(rcu_state.gp_kthread);
203 		WRITE_ONCE(rcu_state.jiffies_kick_kthreads, j + HZ);
204 	}
205 }
206 
207 /*
208  * Handler for the irq_work request posted about halfway into the RCU CPU
209  * stall timeout, and used to detect excessive irq disabling.  Set state
210  * appropriately, but just complain if there is unexpected state on entry.
211  */
212 static void rcu_iw_handler(struct irq_work *iwp)
213 {
214 	struct rcu_data *rdp;
215 	struct rcu_node *rnp;
216 
217 	rdp = container_of(iwp, struct rcu_data, rcu_iw);
218 	rnp = rdp->mynode;
219 	raw_spin_lock_rcu_node(rnp);
220 	if (!WARN_ON_ONCE(!rdp->rcu_iw_pending)) {
221 		rdp->rcu_iw_gp_seq = rnp->gp_seq;
222 		rdp->rcu_iw_pending = false;
223 	}
224 	raw_spin_unlock_rcu_node(rnp);
225 }
226 
227 //////////////////////////////////////////////////////////////////////////////
228 //
229 // Printing RCU CPU stall warnings
230 
231 #ifdef CONFIG_PREEMPT_RCU
232 
233 /*
234  * Dump detailed information for all tasks blocking the current RCU
235  * grace period on the specified rcu_node structure.
236  */
237 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
238 {
239 	unsigned long flags;
240 	struct task_struct *t;
241 
242 	raw_spin_lock_irqsave_rcu_node(rnp, flags);
243 	if (!rcu_preempt_blocked_readers_cgp(rnp)) {
244 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
245 		return;
246 	}
247 	t = list_entry(rnp->gp_tasks->prev,
248 		       struct task_struct, rcu_node_entry);
249 	list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
250 		/*
251 		 * We could be printing a lot while holding a spinlock.
252 		 * Avoid triggering hard lockup.
253 		 */
254 		touch_nmi_watchdog();
255 		sched_show_task(t);
256 	}
257 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
258 }
259 
260 // Communicate task state back to the RCU CPU stall warning request.
261 struct rcu_stall_chk_rdr {
262 	int nesting;
263 	union rcu_special rs;
264 	bool on_blkd_list;
265 };
266 
267 /*
268  * Report out the state of a not-running task that is stalling the
269  * current RCU grace period.
270  */
271 static int check_slow_task(struct task_struct *t, void *arg)
272 {
273 	struct rcu_stall_chk_rdr *rscrp = arg;
274 
275 	if (task_curr(t))
276 		return -EBUSY; // It is running, so decline to inspect it.
277 	rscrp->nesting = t->rcu_read_lock_nesting;
278 	rscrp->rs = t->rcu_read_unlock_special;
279 	rscrp->on_blkd_list = !list_empty(&t->rcu_node_entry);
280 	return 0;
281 }
282 
283 /*
284  * Scan the current list of tasks blocked within RCU read-side critical
285  * sections, printing out the tid of each of the first few of them.
286  */
287 static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags)
288 	__releases(rnp->lock)
289 {
290 	int i = 0;
291 	int ndetected = 0;
292 	struct rcu_stall_chk_rdr rscr;
293 	struct task_struct *t;
294 	struct task_struct *ts[8];
295 
296 	lockdep_assert_irqs_disabled();
297 	if (!rcu_preempt_blocked_readers_cgp(rnp)) {
298 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
299 		return 0;
300 	}
301 	pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
302 	       rnp->level, rnp->grplo, rnp->grphi);
303 	t = list_entry(rnp->gp_tasks->prev,
304 		       struct task_struct, rcu_node_entry);
305 	list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
306 		get_task_struct(t);
307 		ts[i++] = t;
308 		if (i >= ARRAY_SIZE(ts))
309 			break;
310 	}
311 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
312 	while (i) {
313 		t = ts[--i];
314 		if (task_call_func(t, check_slow_task, &rscr))
315 			pr_cont(" P%d", t->pid);
316 		else
317 			pr_cont(" P%d/%d:%c%c%c%c",
318 				t->pid, rscr.nesting,
319 				".b"[rscr.rs.b.blocked],
320 				".q"[rscr.rs.b.need_qs],
321 				".e"[rscr.rs.b.exp_hint],
322 				".l"[rscr.on_blkd_list]);
323 		lockdep_assert_irqs_disabled();
324 		put_task_struct(t);
325 		ndetected++;
326 	}
327 	pr_cont("\n");
328 	return ndetected;
329 }
330 
331 #else /* #ifdef CONFIG_PREEMPT_RCU */
332 
333 /*
334  * Because preemptible RCU does not exist, we never have to check for
335  * tasks blocked within RCU read-side critical sections.
336  */
337 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
338 {
339 }
340 
341 /*
342  * Because preemptible RCU does not exist, we never have to check for
343  * tasks blocked within RCU read-side critical sections.
344  */
345 static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags)
346 	__releases(rnp->lock)
347 {
348 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
349 	return 0;
350 }
351 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
352 
353 /*
354  * Dump stacks of all tasks running on stalled CPUs.  First try using
355  * NMIs, but fall back to manual remote stack tracing on architectures
356  * that don't support NMI-based stack dumps.  The NMI-triggered stack
357  * traces are more accurate because they are printed by the target CPU.
358  */
359 static void rcu_dump_cpu_stacks(void)
360 {
361 	int cpu;
362 	unsigned long flags;
363 	struct rcu_node *rnp;
364 
365 	rcu_for_each_leaf_node(rnp) {
366 		raw_spin_lock_irqsave_rcu_node(rnp, flags);
367 		for_each_leaf_node_possible_cpu(rnp, cpu)
368 			if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
369 				if (cpu_is_offline(cpu))
370 					pr_err("Offline CPU %d blocking current GP.\n", cpu);
371 				else
372 					dump_cpu_task(cpu);
373 			}
374 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
375 	}
376 }
377 
378 static const char * const gp_state_names[] = {
379 	[RCU_GP_IDLE] = "RCU_GP_IDLE",
380 	[RCU_GP_WAIT_GPS] = "RCU_GP_WAIT_GPS",
381 	[RCU_GP_DONE_GPS] = "RCU_GP_DONE_GPS",
382 	[RCU_GP_ONOFF] = "RCU_GP_ONOFF",
383 	[RCU_GP_INIT] = "RCU_GP_INIT",
384 	[RCU_GP_WAIT_FQS] = "RCU_GP_WAIT_FQS",
385 	[RCU_GP_DOING_FQS] = "RCU_GP_DOING_FQS",
386 	[RCU_GP_CLEANUP] = "RCU_GP_CLEANUP",
387 	[RCU_GP_CLEANED] = "RCU_GP_CLEANED",
388 };
389 
390 /*
391  * Convert a ->gp_state value to a character string.
392  */
393 static const char *gp_state_getname(short gs)
394 {
395 	if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
396 		return "???";
397 	return gp_state_names[gs];
398 }
399 
400 /* Is the RCU grace-period kthread being starved of CPU time? */
401 static bool rcu_is_gp_kthread_starving(unsigned long *jp)
402 {
403 	unsigned long j = jiffies - READ_ONCE(rcu_state.gp_activity);
404 
405 	if (jp)
406 		*jp = j;
407 	return j > 2 * HZ;
408 }
409 
410 static bool rcu_is_rcuc_kthread_starving(struct rcu_data *rdp, unsigned long *jp)
411 {
412 	int cpu;
413 	struct task_struct *rcuc;
414 	unsigned long j;
415 
416 	rcuc = rdp->rcu_cpu_kthread_task;
417 	if (!rcuc)
418 		return false;
419 
420 	cpu = task_cpu(rcuc);
421 	if (cpu_is_offline(cpu) || idle_cpu(cpu))
422 		return false;
423 
424 	j = jiffies - READ_ONCE(rdp->rcuc_activity);
425 
426 	if (jp)
427 		*jp = j;
428 	return j > 2 * HZ;
429 }
430 
431 /*
432  * Print out diagnostic information for the specified stalled CPU.
433  *
434  * If the specified CPU is aware of the current RCU grace period, then
435  * print the number of scheduling clock interrupts the CPU has taken
436  * during the time that it has been aware.  Otherwise, print the number
437  * of RCU grace periods that this CPU is ignorant of, for example, "1"
438  * if the CPU was aware of the previous grace period.
439  *
440  * Also print out idle info.
441  */
442 static void print_cpu_stall_info(int cpu)
443 {
444 	unsigned long delta;
445 	bool falsepositive;
446 	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
447 	char *ticks_title;
448 	unsigned long ticks_value;
449 	bool rcuc_starved;
450 	unsigned long j;
451 	char buf[32];
452 
453 	/*
454 	 * We could be printing a lot while holding a spinlock.  Avoid
455 	 * triggering hard lockup.
456 	 */
457 	touch_nmi_watchdog();
458 
459 	ticks_value = rcu_seq_ctr(rcu_state.gp_seq - rdp->gp_seq);
460 	if (ticks_value) {
461 		ticks_title = "GPs behind";
462 	} else {
463 		ticks_title = "ticks this GP";
464 		ticks_value = rdp->ticks_this_gp;
465 	}
466 	delta = rcu_seq_ctr(rdp->mynode->gp_seq - rdp->rcu_iw_gp_seq);
467 	falsepositive = rcu_is_gp_kthread_starving(NULL) &&
468 			rcu_dynticks_in_eqs(rcu_dynticks_snap(cpu));
469 	rcuc_starved = rcu_is_rcuc_kthread_starving(rdp, &j);
470 	if (rcuc_starved)
471 		sprintf(buf, " rcuc=%ld jiffies(starved)", j);
472 	pr_err("\t%d-%c%c%c%c: (%lu %s) idle=%04x/%ld/%#lx softirq=%u/%u fqs=%ld%s%s\n",
473 	       cpu,
474 	       "O."[!!cpu_online(cpu)],
475 	       "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)],
476 	       "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)],
477 	       !IS_ENABLED(CONFIG_IRQ_WORK) ? '?' :
478 			rdp->rcu_iw_pending ? (int)min(delta, 9UL) + '0' :
479 				"!."[!delta],
480 	       ticks_value, ticks_title,
481 	       rcu_dynticks_snap(cpu) & 0xffff,
482 	       ct_dynticks_nesting_cpu(cpu), ct_dynticks_nmi_nesting_cpu(cpu),
483 	       rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
484 	       data_race(rcu_state.n_force_qs) - rcu_state.n_force_qs_gpstart,
485 	       rcuc_starved ? buf : "",
486 	       falsepositive ? " (false positive?)" : "");
487 }
488 
489 /* Complain about starvation of grace-period kthread.  */
490 static void rcu_check_gp_kthread_starvation(void)
491 {
492 	int cpu;
493 	struct task_struct *gpk = rcu_state.gp_kthread;
494 	unsigned long j;
495 
496 	if (rcu_is_gp_kthread_starving(&j)) {
497 		cpu = gpk ? task_cpu(gpk) : -1;
498 		pr_err("%s kthread starved for %ld jiffies! g%ld f%#x %s(%d) ->state=%#x ->cpu=%d\n",
499 		       rcu_state.name, j,
500 		       (long)rcu_seq_current(&rcu_state.gp_seq),
501 		       data_race(READ_ONCE(rcu_state.gp_flags)),
502 		       gp_state_getname(rcu_state.gp_state),
503 		       data_race(READ_ONCE(rcu_state.gp_state)),
504 		       gpk ? data_race(READ_ONCE(gpk->__state)) : ~0, cpu);
505 		if (gpk) {
506 			pr_err("\tUnless %s kthread gets sufficient CPU time, OOM is now expected behavior.\n", rcu_state.name);
507 			pr_err("RCU grace-period kthread stack dump:\n");
508 			sched_show_task(gpk);
509 			if (cpu >= 0) {
510 				if (cpu_is_offline(cpu)) {
511 					pr_err("RCU GP kthread last ran on offline CPU %d.\n", cpu);
512 				} else  {
513 					pr_err("Stack dump where RCU GP kthread last ran:\n");
514 					dump_cpu_task(cpu);
515 				}
516 			}
517 			wake_up_process(gpk);
518 		}
519 	}
520 }
521 
522 /* Complain about missing wakeups from expired fqs wait timer */
523 static void rcu_check_gp_kthread_expired_fqs_timer(void)
524 {
525 	struct task_struct *gpk = rcu_state.gp_kthread;
526 	short gp_state;
527 	unsigned long jiffies_fqs;
528 	int cpu;
529 
530 	/*
531 	 * Order reads of .gp_state and .jiffies_force_qs.
532 	 * Matching smp_wmb() is present in rcu_gp_fqs_loop().
533 	 */
534 	gp_state = smp_load_acquire(&rcu_state.gp_state);
535 	jiffies_fqs = READ_ONCE(rcu_state.jiffies_force_qs);
536 
537 	if (gp_state == RCU_GP_WAIT_FQS &&
538 	    time_after(jiffies, jiffies_fqs + RCU_STALL_MIGHT_MIN) &&
539 	    gpk && !READ_ONCE(gpk->on_rq)) {
540 		cpu = task_cpu(gpk);
541 		pr_err("%s kthread timer wakeup didn't happen for %ld jiffies! g%ld f%#x %s(%d) ->state=%#x\n",
542 		       rcu_state.name, (jiffies - jiffies_fqs),
543 		       (long)rcu_seq_current(&rcu_state.gp_seq),
544 		       data_race(rcu_state.gp_flags),
545 		       gp_state_getname(RCU_GP_WAIT_FQS), RCU_GP_WAIT_FQS,
546 		       data_race(READ_ONCE(gpk->__state)));
547 		pr_err("\tPossible timer handling issue on cpu=%d timer-softirq=%u\n",
548 		       cpu, kstat_softirqs_cpu(TIMER_SOFTIRQ, cpu));
549 	}
550 }
551 
552 static void print_other_cpu_stall(unsigned long gp_seq, unsigned long gps)
553 {
554 	int cpu;
555 	unsigned long flags;
556 	unsigned long gpa;
557 	unsigned long j;
558 	int ndetected = 0;
559 	struct rcu_node *rnp;
560 	long totqlen = 0;
561 
562 	lockdep_assert_irqs_disabled();
563 
564 	/* Kick and suppress, if so configured. */
565 	rcu_stall_kick_kthreads();
566 	if (rcu_stall_is_suppressed())
567 		return;
568 
569 	/*
570 	 * OK, time to rat on our buddy...
571 	 * See Documentation/RCU/stallwarn.rst for info on how to debug
572 	 * RCU CPU stall warnings.
573 	 */
574 	trace_rcu_stall_warning(rcu_state.name, TPS("StallDetected"));
575 	pr_err("INFO: %s detected stalls on CPUs/tasks:\n", rcu_state.name);
576 	rcu_for_each_leaf_node(rnp) {
577 		raw_spin_lock_irqsave_rcu_node(rnp, flags);
578 		if (rnp->qsmask != 0) {
579 			for_each_leaf_node_possible_cpu(rnp, cpu)
580 				if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
581 					print_cpu_stall_info(cpu);
582 					ndetected++;
583 				}
584 		}
585 		ndetected += rcu_print_task_stall(rnp, flags); // Releases rnp->lock.
586 		lockdep_assert_irqs_disabled();
587 	}
588 
589 	for_each_possible_cpu(cpu)
590 		totqlen += rcu_get_n_cbs_cpu(cpu);
591 	pr_cont("\t(detected by %d, t=%ld jiffies, g=%ld, q=%lu ncpus=%d)\n",
592 	       smp_processor_id(), (long)(jiffies - gps),
593 	       (long)rcu_seq_current(&rcu_state.gp_seq), totqlen, rcu_state.n_online_cpus);
594 	if (ndetected) {
595 		rcu_dump_cpu_stacks();
596 
597 		/* Complain about tasks blocking the grace period. */
598 		rcu_for_each_leaf_node(rnp)
599 			rcu_print_detail_task_stall_rnp(rnp);
600 	} else {
601 		if (rcu_seq_current(&rcu_state.gp_seq) != gp_seq) {
602 			pr_err("INFO: Stall ended before state dump start\n");
603 		} else {
604 			j = jiffies;
605 			gpa = data_race(READ_ONCE(rcu_state.gp_activity));
606 			pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
607 			       rcu_state.name, j - gpa, j, gpa,
608 			       data_race(READ_ONCE(jiffies_till_next_fqs)),
609 			       data_race(READ_ONCE(rcu_get_root()->qsmask)));
610 		}
611 	}
612 	/* Rewrite if needed in case of slow consoles. */
613 	if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
614 		WRITE_ONCE(rcu_state.jiffies_stall,
615 			   jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
616 
617 	rcu_check_gp_kthread_expired_fqs_timer();
618 	rcu_check_gp_kthread_starvation();
619 
620 	panic_on_rcu_stall();
621 
622 	rcu_force_quiescent_state();  /* Kick them all. */
623 }
624 
625 static void print_cpu_stall(unsigned long gps)
626 {
627 	int cpu;
628 	unsigned long flags;
629 	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
630 	struct rcu_node *rnp = rcu_get_root();
631 	long totqlen = 0;
632 
633 	lockdep_assert_irqs_disabled();
634 
635 	/* Kick and suppress, if so configured. */
636 	rcu_stall_kick_kthreads();
637 	if (rcu_stall_is_suppressed())
638 		return;
639 
640 	/*
641 	 * OK, time to rat on ourselves...
642 	 * See Documentation/RCU/stallwarn.rst for info on how to debug
643 	 * RCU CPU stall warnings.
644 	 */
645 	trace_rcu_stall_warning(rcu_state.name, TPS("SelfDetected"));
646 	pr_err("INFO: %s self-detected stall on CPU\n", rcu_state.name);
647 	raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags);
648 	print_cpu_stall_info(smp_processor_id());
649 	raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags);
650 	for_each_possible_cpu(cpu)
651 		totqlen += rcu_get_n_cbs_cpu(cpu);
652 	pr_cont("\t(t=%lu jiffies g=%ld q=%lu ncpus=%d)\n",
653 		jiffies - gps,
654 		(long)rcu_seq_current(&rcu_state.gp_seq), totqlen, rcu_state.n_online_cpus);
655 
656 	rcu_check_gp_kthread_expired_fqs_timer();
657 	rcu_check_gp_kthread_starvation();
658 
659 	rcu_dump_cpu_stacks();
660 
661 	raw_spin_lock_irqsave_rcu_node(rnp, flags);
662 	/* Rewrite if needed in case of slow consoles. */
663 	if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
664 		WRITE_ONCE(rcu_state.jiffies_stall,
665 			   jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
666 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
667 
668 	panic_on_rcu_stall();
669 
670 	/*
671 	 * Attempt to revive the RCU machinery by forcing a context switch.
672 	 *
673 	 * A context switch would normally allow the RCU state machine to make
674 	 * progress and it could be we're stuck in kernel space without context
675 	 * switches for an entirely unreasonable amount of time.
676 	 */
677 	set_tsk_need_resched(current);
678 	set_preempt_need_resched();
679 }
680 
681 static void check_cpu_stall(struct rcu_data *rdp)
682 {
683 	bool didstall = false;
684 	unsigned long gs1;
685 	unsigned long gs2;
686 	unsigned long gps;
687 	unsigned long j;
688 	unsigned long jn;
689 	unsigned long js;
690 	struct rcu_node *rnp;
691 
692 	lockdep_assert_irqs_disabled();
693 	if ((rcu_stall_is_suppressed() && !READ_ONCE(rcu_kick_kthreads)) ||
694 	    !rcu_gp_in_progress())
695 		return;
696 	rcu_stall_kick_kthreads();
697 	j = jiffies;
698 
699 	/*
700 	 * Lots of memory barriers to reject false positives.
701 	 *
702 	 * The idea is to pick up rcu_state.gp_seq, then
703 	 * rcu_state.jiffies_stall, then rcu_state.gp_start, and finally
704 	 * another copy of rcu_state.gp_seq.  These values are updated in
705 	 * the opposite order with memory barriers (or equivalent) during
706 	 * grace-period initialization and cleanup.  Now, a false positive
707 	 * can occur if we get an new value of rcu_state.gp_start and a old
708 	 * value of rcu_state.jiffies_stall.  But given the memory barriers,
709 	 * the only way that this can happen is if one grace period ends
710 	 * and another starts between these two fetches.  This is detected
711 	 * by comparing the second fetch of rcu_state.gp_seq with the
712 	 * previous fetch from rcu_state.gp_seq.
713 	 *
714 	 * Given this check, comparisons of jiffies, rcu_state.jiffies_stall,
715 	 * and rcu_state.gp_start suffice to forestall false positives.
716 	 */
717 	gs1 = READ_ONCE(rcu_state.gp_seq);
718 	smp_rmb(); /* Pick up ->gp_seq first... */
719 	js = READ_ONCE(rcu_state.jiffies_stall);
720 	smp_rmb(); /* ...then ->jiffies_stall before the rest... */
721 	gps = READ_ONCE(rcu_state.gp_start);
722 	smp_rmb(); /* ...and finally ->gp_start before ->gp_seq again. */
723 	gs2 = READ_ONCE(rcu_state.gp_seq);
724 	if (gs1 != gs2 ||
725 	    ULONG_CMP_LT(j, js) ||
726 	    ULONG_CMP_GE(gps, js))
727 		return; /* No stall or GP completed since entering function. */
728 	rnp = rdp->mynode;
729 	jn = jiffies + ULONG_MAX / 2;
730 	if (rcu_gp_in_progress() &&
731 	    (READ_ONCE(rnp->qsmask) & rdp->grpmask) &&
732 	    cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
733 
734 		/*
735 		 * If a virtual machine is stopped by the host it can look to
736 		 * the watchdog like an RCU stall. Check to see if the host
737 		 * stopped the vm.
738 		 */
739 		if (kvm_check_and_clear_guest_paused())
740 			return;
741 
742 		/* We haven't checked in, so go dump stack. */
743 		print_cpu_stall(gps);
744 		if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
745 			rcu_ftrace_dump(DUMP_ALL);
746 		didstall = true;
747 
748 	} else if (rcu_gp_in_progress() &&
749 		   ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY) &&
750 		   cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
751 
752 		/*
753 		 * If a virtual machine is stopped by the host it can look to
754 		 * the watchdog like an RCU stall. Check to see if the host
755 		 * stopped the vm.
756 		 */
757 		if (kvm_check_and_clear_guest_paused())
758 			return;
759 
760 		/* They had a few time units to dump stack, so complain. */
761 		print_other_cpu_stall(gs2, gps);
762 		if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
763 			rcu_ftrace_dump(DUMP_ALL);
764 		didstall = true;
765 	}
766 	if (didstall && READ_ONCE(rcu_state.jiffies_stall) == jn) {
767 		jn = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
768 		WRITE_ONCE(rcu_state.jiffies_stall, jn);
769 	}
770 }
771 
772 //////////////////////////////////////////////////////////////////////////////
773 //
774 // RCU forward-progress mechanisms, including of callback invocation.
775 
776 
777 /*
778  * Check to see if a failure to end RCU priority inversion was due to
779  * a CPU not passing through a quiescent state.  When this happens, there
780  * is nothing that RCU priority boosting can do to help, so we shouldn't
781  * count this as an RCU priority boosting failure.  A return of true says
782  * RCU priority boosting is to blame, and false says otherwise.  If false
783  * is returned, the first of the CPUs to blame is stored through cpup.
784  * If there was no CPU blocking the current grace period, but also nothing
785  * in need of being boosted, *cpup is set to -1.  This can happen in case
786  * of vCPU preemption while the last CPU is reporting its quiscent state,
787  * for example.
788  *
789  * If cpup is NULL, then a lockless quick check is carried out, suitable
790  * for high-rate usage.  On the other hand, if cpup is non-NULL, each
791  * rcu_node structure's ->lock is acquired, ruling out high-rate usage.
792  */
793 bool rcu_check_boost_fail(unsigned long gp_state, int *cpup)
794 {
795 	bool atb = false;
796 	int cpu;
797 	unsigned long flags;
798 	struct rcu_node *rnp;
799 
800 	rcu_for_each_leaf_node(rnp) {
801 		if (!cpup) {
802 			if (data_race(READ_ONCE(rnp->qsmask))) {
803 				return false;
804 			} else {
805 				if (READ_ONCE(rnp->gp_tasks))
806 					atb = true;
807 				continue;
808 			}
809 		}
810 		*cpup = -1;
811 		raw_spin_lock_irqsave_rcu_node(rnp, flags);
812 		if (rnp->gp_tasks)
813 			atb = true;
814 		if (!rnp->qsmask) {
815 			// No CPUs without quiescent states for this rnp.
816 			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
817 			continue;
818 		}
819 		// Find the first holdout CPU.
820 		for_each_leaf_node_possible_cpu(rnp, cpu) {
821 			if (rnp->qsmask & (1UL << (cpu - rnp->grplo))) {
822 				raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
823 				*cpup = cpu;
824 				return false;
825 			}
826 		}
827 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
828 	}
829 	// Can't blame CPUs, so must blame RCU priority boosting.
830 	return atb;
831 }
832 EXPORT_SYMBOL_GPL(rcu_check_boost_fail);
833 
834 /*
835  * Show the state of the grace-period kthreads.
836  */
837 void show_rcu_gp_kthreads(void)
838 {
839 	unsigned long cbs = 0;
840 	int cpu;
841 	unsigned long j;
842 	unsigned long ja;
843 	unsigned long jr;
844 	unsigned long js;
845 	unsigned long jw;
846 	struct rcu_data *rdp;
847 	struct rcu_node *rnp;
848 	struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
849 
850 	j = jiffies;
851 	ja = j - data_race(READ_ONCE(rcu_state.gp_activity));
852 	jr = j - data_race(READ_ONCE(rcu_state.gp_req_activity));
853 	js = j - data_race(READ_ONCE(rcu_state.gp_start));
854 	jw = j - data_race(READ_ONCE(rcu_state.gp_wake_time));
855 	pr_info("%s: wait state: %s(%d) ->state: %#x ->rt_priority %u delta ->gp_start %lu ->gp_activity %lu ->gp_req_activity %lu ->gp_wake_time %lu ->gp_wake_seq %ld ->gp_seq %ld ->gp_seq_needed %ld ->gp_max %lu ->gp_flags %#x\n",
856 		rcu_state.name, gp_state_getname(rcu_state.gp_state),
857 		data_race(READ_ONCE(rcu_state.gp_state)),
858 		t ? data_race(READ_ONCE(t->__state)) : 0x1ffff, t ? t->rt_priority : 0xffU,
859 		js, ja, jr, jw, (long)data_race(READ_ONCE(rcu_state.gp_wake_seq)),
860 		(long)data_race(READ_ONCE(rcu_state.gp_seq)),
861 		(long)data_race(READ_ONCE(rcu_get_root()->gp_seq_needed)),
862 		data_race(READ_ONCE(rcu_state.gp_max)),
863 		data_race(READ_ONCE(rcu_state.gp_flags)));
864 	rcu_for_each_node_breadth_first(rnp) {
865 		if (ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq), READ_ONCE(rnp->gp_seq_needed)) &&
866 		    !data_race(READ_ONCE(rnp->qsmask)) && !data_race(READ_ONCE(rnp->boost_tasks)) &&
867 		    !data_race(READ_ONCE(rnp->exp_tasks)) && !data_race(READ_ONCE(rnp->gp_tasks)))
868 			continue;
869 		pr_info("\trcu_node %d:%d ->gp_seq %ld ->gp_seq_needed %ld ->qsmask %#lx %c%c%c%c ->n_boosts %ld\n",
870 			rnp->grplo, rnp->grphi,
871 			(long)data_race(READ_ONCE(rnp->gp_seq)),
872 			(long)data_race(READ_ONCE(rnp->gp_seq_needed)),
873 			data_race(READ_ONCE(rnp->qsmask)),
874 			".b"[!!data_race(READ_ONCE(rnp->boost_kthread_task))],
875 			".B"[!!data_race(READ_ONCE(rnp->boost_tasks))],
876 			".E"[!!data_race(READ_ONCE(rnp->exp_tasks))],
877 			".G"[!!data_race(READ_ONCE(rnp->gp_tasks))],
878 			data_race(READ_ONCE(rnp->n_boosts)));
879 		if (!rcu_is_leaf_node(rnp))
880 			continue;
881 		for_each_leaf_node_possible_cpu(rnp, cpu) {
882 			rdp = per_cpu_ptr(&rcu_data, cpu);
883 			if (READ_ONCE(rdp->gpwrap) ||
884 			    ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq),
885 					 READ_ONCE(rdp->gp_seq_needed)))
886 				continue;
887 			pr_info("\tcpu %d ->gp_seq_needed %ld\n",
888 				cpu, (long)data_race(READ_ONCE(rdp->gp_seq_needed)));
889 		}
890 	}
891 	for_each_possible_cpu(cpu) {
892 		rdp = per_cpu_ptr(&rcu_data, cpu);
893 		cbs += data_race(READ_ONCE(rdp->n_cbs_invoked));
894 		if (rcu_segcblist_is_offloaded(&rdp->cblist))
895 			show_rcu_nocb_state(rdp);
896 	}
897 	pr_info("RCU callbacks invoked since boot: %lu\n", cbs);
898 	show_rcu_tasks_gp_kthreads();
899 }
900 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
901 
902 /*
903  * This function checks for grace-period requests that fail to motivate
904  * RCU to come out of its idle mode.
905  */
906 static void rcu_check_gp_start_stall(struct rcu_node *rnp, struct rcu_data *rdp,
907 				     const unsigned long gpssdelay)
908 {
909 	unsigned long flags;
910 	unsigned long j;
911 	struct rcu_node *rnp_root = rcu_get_root();
912 	static atomic_t warned = ATOMIC_INIT(0);
913 
914 	if (!IS_ENABLED(CONFIG_PROVE_RCU) || rcu_gp_in_progress() ||
915 	    ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
916 			 READ_ONCE(rnp_root->gp_seq_needed)) ||
917 	    !smp_load_acquire(&rcu_state.gp_kthread)) // Get stable kthread.
918 		return;
919 	j = jiffies; /* Expensive access, and in common case don't get here. */
920 	if (time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
921 	    time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
922 	    atomic_read(&warned))
923 		return;
924 
925 	raw_spin_lock_irqsave_rcu_node(rnp, flags);
926 	j = jiffies;
927 	if (rcu_gp_in_progress() ||
928 	    ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
929 			 READ_ONCE(rnp_root->gp_seq_needed)) ||
930 	    time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
931 	    time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
932 	    atomic_read(&warned)) {
933 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
934 		return;
935 	}
936 	/* Hold onto the leaf lock to make others see warned==1. */
937 
938 	if (rnp_root != rnp)
939 		raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
940 	j = jiffies;
941 	if (rcu_gp_in_progress() ||
942 	    ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
943 			 READ_ONCE(rnp_root->gp_seq_needed)) ||
944 	    time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
945 	    time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
946 	    atomic_xchg(&warned, 1)) {
947 		if (rnp_root != rnp)
948 			/* irqs remain disabled. */
949 			raw_spin_unlock_rcu_node(rnp_root);
950 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
951 		return;
952 	}
953 	WARN_ON(1);
954 	if (rnp_root != rnp)
955 		raw_spin_unlock_rcu_node(rnp_root);
956 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
957 	show_rcu_gp_kthreads();
958 }
959 
960 /*
961  * Do a forward-progress check for rcutorture.  This is normally invoked
962  * due to an OOM event.  The argument "j" gives the time period during
963  * which rcutorture would like progress to have been made.
964  */
965 void rcu_fwd_progress_check(unsigned long j)
966 {
967 	unsigned long cbs;
968 	int cpu;
969 	unsigned long max_cbs = 0;
970 	int max_cpu = -1;
971 	struct rcu_data *rdp;
972 
973 	if (rcu_gp_in_progress()) {
974 		pr_info("%s: GP age %lu jiffies\n",
975 			__func__, jiffies - data_race(READ_ONCE(rcu_state.gp_start)));
976 		show_rcu_gp_kthreads();
977 	} else {
978 		pr_info("%s: Last GP end %lu jiffies ago\n",
979 			__func__, jiffies - data_race(READ_ONCE(rcu_state.gp_end)));
980 		preempt_disable();
981 		rdp = this_cpu_ptr(&rcu_data);
982 		rcu_check_gp_start_stall(rdp->mynode, rdp, j);
983 		preempt_enable();
984 	}
985 	for_each_possible_cpu(cpu) {
986 		cbs = rcu_get_n_cbs_cpu(cpu);
987 		if (!cbs)
988 			continue;
989 		if (max_cpu < 0)
990 			pr_info("%s: callbacks", __func__);
991 		pr_cont(" %d: %lu", cpu, cbs);
992 		if (cbs <= max_cbs)
993 			continue;
994 		max_cbs = cbs;
995 		max_cpu = cpu;
996 	}
997 	if (max_cpu >= 0)
998 		pr_cont("\n");
999 }
1000 EXPORT_SYMBOL_GPL(rcu_fwd_progress_check);
1001 
1002 /* Commandeer a sysrq key to dump RCU's tree. */
1003 static bool sysrq_rcu;
1004 module_param(sysrq_rcu, bool, 0444);
1005 
1006 /* Dump grace-period-request information due to commandeered sysrq. */
1007 static void sysrq_show_rcu(int key)
1008 {
1009 	show_rcu_gp_kthreads();
1010 }
1011 
1012 static const struct sysrq_key_op sysrq_rcudump_op = {
1013 	.handler = sysrq_show_rcu,
1014 	.help_msg = "show-rcu(y)",
1015 	.action_msg = "Show RCU tree",
1016 	.enable_mask = SYSRQ_ENABLE_DUMP,
1017 };
1018 
1019 static int __init rcu_sysrq_init(void)
1020 {
1021 	if (sysrq_rcu)
1022 		return register_sysrq_key('y', &sysrq_rcudump_op);
1023 	return 0;
1024 }
1025 early_initcall(rcu_sysrq_init);
1026